Skin lesion classification via hybrid localization-explainability approaches

Tez Türü: Doktora

Tezin Yürütüldüğü Kurum: Kocaeli Üniversitesi, Fen Bilimleri Enstitüsü, Makatronik Mühendisliği Bölümü, Türkiye

Tez Danışmanı: Prof. Dr. Hüseyin Metin Ertunç

Tezin Onay Tarihi: 2025

Tezin Dili: İngilizce

Özet:

Skin cancer constitutes a significant portion of dermatological malignancies globally, with its incidence continuing to rise due to environmental and lifestyle factors. Because delayed diagnosis is strongly associated with adverse clinical outcomes, the development of accurate, early, and objective diagnostic tools is of paramount importance. However, conventional dermoscopic evaluation remains prone to inter-observer variability and often falls short when confronting complex lesion morphologies. This thesis addresses these critical challenges by designing and systematically evaluating two explainable, end-to-end deep learning frameworks for multiclass skin cancer classification from dermoscopic images. The first framework integrates a YOLOv8-based lesion localization module for precise region-of-interest detection, followed by the fine-tuning of eight pretrained vision transformer and convolution-based architectures on both lesion-focused and full-image inputs. The integration of lesion localization resulted in a notable absolute improvement of approximately 4.0% in classification accuracy, with the ViT+YOLOv8 hybrid pipeline achieving 93.07% on the HAM10000 dataset, surpassing all tested baselines. The second framework explores six novel depth-wise and channel-wise fusion architectures that integrate feature representations from DenseNet201, ResNet50V2, and Vision Transformer models. These fusion models are examined for their complementary strengths, with some achieving superior classification performance (e.g., FM6 at 90.57%), while others offer enhanced explainability. Both frameworks incorporate Grad-CAM and SHAP-based interpretability mechanisms to support clinical validation and trust. Overall, this research contributes a comprehensive, explainability-aware investigation into localization and fusion strategies, offering clinically relevant, generalizable AI systems that address key limitations in current automated skin cancer diagnostics and support the development of trustworthy medical decision-support tools.